Parametric vs Non-Parametric Test: How to Choose the Right One for Your Data

Parametric vs non-parametric test decisions come down to three things: your data type, your sample size, and whether your data meets the normality assumption. I’ve spent the last 12 years helping dissertation students make this exact call, usually two weeks before their submission deadline, and I can tell you it’s simpler than the textbooks make it look.

This guide gives you the exact framework I use with my own clients. No jargon, no 40-page stats manual required.

What Is a Parametric Test?

A parametric test assumes your data follows a known distribution, usually the normal distribution, and it uses that assumption to make inferences about the mean. Think t-test, ANOVA, Pearson correlation. If your data genuinely behaves the way the test expects, parametric tests give you more statistical power for the same sample size.

That last part matters more than most guides admit. Power means you’re less likely to miss a real effect that actually exists in your data. Throwing that away for no good reason is a mistake I see constantly.

What Is a Non-Parametric Test?

A non-parametric test makes no assumption about the underlying distribution of your data. Instead of working with means, it works with ranks or medians. Mann-Whitney U, Kruskal-Wallis, Spearman correlation, and Wilcoxon signed-rank all fall in this bucket.

This is the definition worth remembering: a parametric test estimates population parameters (mean, variance) assuming a known distribution, while a non-parametric test makes inferences without that assumption, using ranks instead of raw values.

Parametric vs Non-Parametric Test: The Core Differences

Aspect	Parametric Test	Non-Parametric Test
Assumption	Normal distribution	No distribution assumed
Central measure	Mean	Median
Data type needed	Interval or ratio	Ordinal, nominal, or non-normal continuous
Statistical power	Higher, when assumptions hold	Lower, generally
Sensitivity to outliers	High	Low
Sample size comfort	Needs a reasonable size for the mean to stabilise	Works with small samples too

The 3-Step Framework I Use to Choose the Right Test

I don’t look at a flowchart with fifteen branches. Choosing the right statistical test comes down to three questions, asked in order.

1. What type of data do you have? Nominal or ordinal data (rankings, categories, Likert scales) almost always points to a non-parametric test. Continuous interval or ratio data keeps the parametric door open.
2. What is your sample size? With very small samples, normality is nearly impossible to confirm, so lean non-parametric. With larger samples, the Central Limit Theorem does a lot of the heavy lifting, and moderate non-normality stops mattering as much.
3. Does your data pass the normality assumption check? Run a Shapiro-Wilk test and look at a Q-Q plot alongside it. Don’t rely on either one alone.

A quick word of caution here, because I see this mistake constantly. Some guides suggest that eyeballing a histogram for symmetry is proof enough of normality. It isn’t, especially with a dissertation-sized sample under 50, where a histogram can look lumpy or skewed purely from random noise, not because the underlying population is actually non-normal. Pair the visual with an actual test result.

Shapiro-Wilk or Kolmogorov-Smirnov? Both check normality, but they’re not interchangeable. Shapiro-Wilk is the sharper tool for smaller samples (under roughly 50), which covers most dissertation datasets. Kolmogorov-Smirnov is better suited to larger samples and to comparing your data against distributions other than the normal one. For a typical thesis with 30 to 150 respondents, I default to Shapiro-Wilk.

If you’d rather not do this manually every time, I built a statistical test selection calculator that walks through exactly these three questions and gives you an answer in under a minute.

Case in point (names and identifying details changed; this reflects a pattern I see repeatedly in client work, not a verbatim case file): an MBA student researching consumer trust in mobile banking had a sample of 45 respondents on a 5-point Likert scale. She wanted to run a t-test because “that’s what everyone in her department used.” Likert data is ordinal. We ran Mann-Whitney U instead, and her results held up fine at the viva.

When to Use Non-Parametric Tests: A Checklist

Reach for a non-parametric test when any of these apply:

• Your data is ordinal or ranked, not truly continuous
• Your sample size is small (under 20 to 25 per group) and normality can’t be confirmed
• Your data has clear outliers you cannot justify removing
• A histogram or Q-Q plot shows obvious skew, not just minor wobble
• You’re measuring something with a floor or ceiling effect (a lot of “0” or maximum-score responses)

The “Always Go Non-Parametric to Be Safe” Myth

I’ve read plenty of stats guides that quietly repeat some version of “when in doubt, just use non-parametric, it’s safer.” I disagree with this advice, and I’ll tell you exactly where it breaks down.

Non-parametric tests are not free of assumptions either. Mann-Whitney U still assumes similarly shaped distributions across your groups if you want to interpret it as comparing medians. Kruskal-Wallis assumes equal variance of ranks. Calling them “assumption-free” is a marketing line, not a statistical fact.

There’s also a second claim I want to push back on: that crossing a sample size of 30 means you can stop worrying about normality entirely, because the Central Limit Theorem takes over. That’s an oversimplification. Convergence to normal depends heavily on how skewed your original data is. A heavily skewed variable, like income data or response times with a floor effect, can need a sample well beyond 30, sometimes into the hundreds, before that safety net actually holds.

More importantly, defaulting to non-parametric when your data is genuinely fine wastes statistical power. In a clinical or behavioural study, that lost power can mean the difference between detecting a real effect and reporting a null result that isn’t really null. Choosing the right statistical test isn’t about picking whichever one feels “safer.” It’s about matching the test to what your data actually looks like.

Parametric vs Non-Parametric Test Equivalents (Cheat Sheet)

Once you know your data type, sample size, and normality result, use this table to pick the exact test.

Design	Parametric Test	Non-Parametric Equivalent
One sample vs known value	One-sample t-test	Wilcoxon signed-rank test
Two independent groups	Independent-samples t-test	Mann-Whitney U test
Two paired groups	Paired-samples t-test	Wilcoxon signed-rank test
3+ independent groups	One-way ANOVA	Kruskal-Wallis test
3+ repeated measures	Repeated-measures ANOVA	Friedman test
Correlation between two continuous variables	Pearson correlation	Spearman rank correlation

t-test vs Mann-Whitney U

Use the independent-samples t-test when comparing means of two independent, normally distributed groups. Switch to Mann-Whitney U when your data is skewed, ordinal, or the sample per group is small.

ANOVA vs Kruskal-Wallis

One-way ANOVA compares means across three or more independent groups, checking equal variance across groups before you trust the result. When normality or equal variance fails, Kruskal-Wallis compares mean ranks instead.

Reading Kruskal-Wallis output trips up a lot of students, so here’s the short version. You’ll see an H-statistic, degrees of freedom, and a p-value, similar layout to ANOVA’s F-statistic. A significant p-value tells you the groups differ in central tendency (the typical or middle value in each group), but it doesn’t tell you which pairs differ; for that, you need a post-hoc test like Dunn’s.

Paired t-test vs Wilcoxon Signed-Rank

Both compare the same subjects measured twice (before and after, pre and post). Go parametric if the differences between pairs are roughly normal, non-parametric if they’re skewed or you have a handful of extreme cases. If your design has more than two time points on the same subjects, you’re actually choosing between paired t-test and repeated-measures ANOVA, which is a related but separate decision.

Pearson vs Spearman Correlation

Pearson measures linear association between two continuous, normally distributed variables. Spearman uses ranks and captures monotonic relationships (meaning that as one variable rises, the other consistently rises or falls, even if not in a straight line), which makes it the safer default for Likert-scale or ordinal survey data.

Which Software Should You Use to Run These Tests?

The test you choose matters more than the software, but the software decides how painful running it will be.

• SPSS is what most dissertation committees expect, and it has menu-driven options for every test in the cheat sheet above, including a built-in Shapiro-Wilk normality check under Explore.
• R gives you more control and is free, useful if your university doesn’t provide an SPSS licence after graduation.
• Stata shows up often in economics and public health dissertations; the syntax for Mann-Whitney (ranksum) and Kruskal-Wallis (kwallis) is short.
• Jamovi and JASP are free, open-source, and built to look like SPSS, a solid option if cost is the only thing stopping you from using proper stats software.

Whatever tool you pick, run the normality check first, then let that result decide the test, not the other way around.

How to Report Your Results (APA Style)

This is the step students rush, and examiners notice. A results section that just states “we used Mann-Whitney U” without the actual statistic reads as incomplete.

For a t-test, report the t-value, degrees of freedom, p-value, and effect size: t(48) = 2.31, p = .025. For Mann-Whitney U, report the U statistic, z-value if your software gives one, and the p-value. The APA’s own numbers and statistics guide and Purdue’s Online Writing Lab both lay out the exact decimal and formatting rules. I keep both bookmarked for my own client work, because SPSS output doesn’t format itself into APA style.

What If Your Supervisor Questions Your Test Choice?

This happens more than students expect, and it used to catch my early clients off guard. Supervisors sometimes push back not because your choice is wrong, but because they were trained on an older or stricter rule of thumb.

Here’s what actually works in that conversation:

• Show your normality check output (Shapiro-Wilk result and Q-Q plot), don’t just state a conclusion
• Explain your sample size reasoning in plain terms, referencing the Central Limit Theorem where relevant
• If you switched from parametric to non-parametric (or vice versa) after seeing the data, say so honestly rather than pretending you planned it from the start

Case in point (details changed to protect client identity): a PhD candidate I mentored ran a Kruskal-Wallis test on a sample of 180, and her supervisor asked why she “didn’t just use ANOVA since the sample was large.” Her Q-Q plot showed a clear bimodal pattern, not just non-normality, meaning the mean wasn’t a meaningful measure of central tendency at all. That one plot ended the argument in the same meeting.

Common Mistakes Dissertation Students Make Choosing a Test

• Picking the test after peeking at whether the p-value “looks better”
• Assuming a large sample fixes every distribution problem, when severe skew or multimodal data still causes issues
• Running a normality test on the raw outcome instead of on residuals, where it actually matters for regression-style analyses
• Treating Likert-scale data as interval by default, without checking whether that assumption is defensible for their specific study
• Never checking equal variance across groups before choosing between standard ANOVA and its Welch alternative

FAQ

---

Still unsure which test fits your data, or wondering if you need an extra pair of expert eyes on your analysis chapter? I go through this exact decision with students every week. You can book a session with our statistics tutoring team and we’ll sort it out together, before your supervisor asks the question for you.

Written by Siddharth Gupta, statistics consultant and dissertation methodology expert with 12+ years of analytics and research guidance experience. Connect on LinkedIn.